兄弟，在数据集上进行了验证没？stn加与不加的性能对比

[ReverseSystem001]

兄弟，在数据集上进行了验证没？stn加与不加的性能对比，麻烦告诉一下吧

[Quillbolt]

When working on this project, I got limited by a very small dataset (around 200-300 samples) so I tried a heavy generated dataset to train the model on (70 - 85%) with varying results. The best result I got is 0.9 for text detection and 0.65 for text recognition

[ReverseSystem001]

恩恩，我这边对stn进行随机初始化，crnn加载预训练模型，训练到3个epoch的时候，loss从1直接升到了74. 你训练的时候加载预训练模型了吗？

[Quillbolt]

I think my loss function is coded in a very bad way to fit the dataset and device I was running(GTX 960) and also try to have it similar to warp-ctc-loss. The vanilla pytorch ctc loss work better in most case. The pre-trained model from https://github.com/meijieru/crnn.pytorch is what I go for as baseline CRNN

[ReverseSystem001]

I think my loss function is coded in a very bad way to fit the dataset and device I was running(GTX 960) and also try to have it similar to warp-ctc-loss. The vanilla pytorch ctc loss work better in most case. The pre-trained model from https://github.com/meijieru/crnn.pytorch is what I go for as baseline CRNN

After adding STN, the STN+CRNN+CTC training process does not converge after 2 epoch. as i said below. During the training, loss rose directly from 1 to 74. Have you ever encountered this situation during training

[Quillbolt]

Yes, I encounter it during training too and here is the link to the problem https://discuss.pytorch.org/t/proper-way-to-use-torch-nn-ctcloss/28909 It is easy for the model output to become the same tensor with ctc_loss